Does everything (process/train/valid/predict) have to be done in the 9 hour kernel?
\nOr can people train offline and import weights ?
", "messages": [ { "id": "988138", "postDate": "08/27/2020 20:08:53", "content": "Does everything (process/train/valid/predict) have to be done in the 9 hour kernel?
\nOr can people train offline and import weights ?
", "rawMarkdown": "Does everything (process/train/valid/predict) have to be done in the 9 hour kernel? \n\nOr can people train offline and import weights ?", "votes": null }, { "id": "988147", "postDate": "08/27/2020 20:19:36", "content": "I believe you can train offline, but include the seperate notebook in the current one!
", "rawMarkdown": "I believe you can train offline, but include the seperate notebook in the current one!", "votes": null }, { "id": "988336", "postDate": "08/28/2020 01:37:32", "content": "@jmaslek The requirement is that GPU notebook should have run time less than 9 hours and you can use TPU for training and TPU notebook should have runtime less than 3 hours. You can train and import weights. This is my understanding. Probably you can also weight for host comment on this.
", "rawMarkdown": "jmaslek The requirement is that GPU notebook should have run time less than 9 hours and you can use TPU for training and TPU notebook should have runtime less than 3 hours. You can train and import weights. This is my understanding. Probably you can also weight for host comment on this.", "votes": null }, { "id": "988654", "postDate": "08/28/2020 07:23:34", "content": "I would love to hear a word from the hosts about submissions such as: https://www.kaggle.com/corochann/save-your-time-submit-without-kernel-inference
\nIt's probably not allowed as a valid submission but it raises the question of how much we are allowed to pre-compute.
", "rawMarkdown": "I would love to hear a word from the hosts about submissions such as: https://www.kaggle.com/corochann/save-your-time-submit-without-kernel-inference\n\nIt's probably not allowed as a valid submission but it raises the question of how much we are allowed to pre-compute.", "votes": null }, { "id": "988952", "postDate": "08/28/2020 12:40:39", "content": "My understanding is that you could process/train/valid on your own machine and submit results by inference kernel on kaggle.
", "rawMarkdown": "My understanding is that you could process/train/valid on your own machine and submit results by inference kernel on kaggle.", "votes": null }, { "id": "989038", "postDate": "08/28/2020 14:01:08", "content": "That seems to make sense. I dont have any fancy offline GPUs to train on, so I was hoping I would have a shot to be competitive given the dataset size and time restrictions.
", "rawMarkdown": "That seems to make sense. I dont have any fancy offline GPUs to train on, so I was hoping I would have a shot to be competitive given the dataset size and time restrictions.", "votes": null }, { "id": "989502", "postDate": "08/28/2020 21:12:16", "content": "You can train and perform inference offline and upload only the resulting CSV file to the inference kernel.
", "rawMarkdown": "You can train and perform inference offline and upload only the resulting CSV file to the inference kernel.", "votes": null }, { "id": "989915", "postDate": "08/29/2020 07:59:43", "content": "Thank you very much for this clarification @iglovikov
\nYour quick and clear answers are highly appreciated.
So, uploading a CSV qualifies for a valid submission (valid in terms of qualification for prizes) and it's not a standard Code Competition.
\nedit: further clarification from host (see below): to be eligble for prize money, you need to provide the code.
I thought the submission requires the model (for rerunning) but not the actual CSV file. That is a lot different from what I thought before.
", "rawMarkdown": "I thought the submission requires the model (for rerunning) but not the actual CSV file. That is a lot different from what I thought before.", "votes": null }, { "id": "993597", "postDate": "09/01/2020 03:24:31", "content": "@ilu000, I believe your code will still be needed for verification when you win a price or medal. @iglovikov may clarify this for us.
", "rawMarkdown": "ilu000, I believe your code will still be needed for verification when you win a price or medal. @iglovikov may clarify this for us.", "votes": null }, { "id": "1007788", "postDate": "09/12/2020 13:55:43", "content": "Yep. If the team is \"in the money\" and are not able to provide the code that reproduces your submission => team will be disqualified.
\nIt is not the case for medal solutions. We will not check solutions that are \"in medals\" but not \"in the money\".
\nIn the last year Lyft's competition second-place team refused to provide the code => they were kicked from the LB. The person who was at the 4th place become third and he provided the code.
", "rawMarkdown": "Yep. If the team is \"in the money\" and are not able to provide the code that reproduces your submission => team will be disqualified. \n\nIt is not the case for medal solutions. We will not check solutions that are \"in medals\" but not \"in the money\".\n\nIn the last year Lyft's competition second-place team refused to provide the code => they were kicked from the LB. The person who was at the 4th place become third and he provided the code.", "votes": null }, { "id": "1008240", "postDate": "09/12/2020 21:43:49", "content": "Thanks for answering @iglovikov.
", "rawMarkdown": "Thanks for answering @iglovikov.", "votes": null } ], "comments": [ { "id": 988147, "author_name": "sshikamaru", "author_url": "", "post_date": "08/27/2020 20:19:36", "content": "I believe you can train offline, but include the seperate notebook in the current one!
", "votes": null, "replies": [] }, { "id": 988336, "author_name": "deepakrajpurushothaman", "author_url": "", "post_date": "08/28/2020 01:37:32", "content": "@jmaslek The requirement is that GPU notebook should have run time less than 9 hours and you can use TPU for training and TPU notebook should have runtime less than 3 hours. You can train and import weights. This is my understanding. Probably you can also weight for host comment on this.
", "votes": null, "replies": [ { "id": 989038, "author_name": "jmaslek", "author_url": "", "post_date": "08/28/2020 14:01:08", "content": "That seems to make sense. I dont have any fancy offline GPUs to train on, so I was hoping I would have a shot to be competitive given the dataset size and time restrictions.
", "votes": null, "replies": [] } ] }, { "id": 988654, "author_name": "ilu000", "author_url": "", "post_date": "08/28/2020 07:23:34", "content": "I would love to hear a word from the hosts about submissions such as: https://www.kaggle.com/corochann/save-your-time-submit-without-kernel-inference
\nIt's probably not allowed as a valid submission but it raises the question of how much we are allowed to pre-compute.
", "votes": null, "replies": [] }, { "id": 988952, "author_name": "wangstone", "author_url": "", "post_date": "08/28/2020 12:40:39", "content": "My understanding is that you could process/train/valid on your own machine and submit results by inference kernel on kaggle.
", "votes": null, "replies": [] }, { "id": 989502, "author_name": "iglovikov", "author_url": "", "post_date": "08/28/2020 21:12:16", "content": "You can train and perform inference offline and upload only the resulting CSV file to the inference kernel.
", "votes": null, "replies": [ { "id": 989915, "author_name": "ilu000", "author_url": "", "post_date": "08/29/2020 07:59:43", "content": "Thank you very much for this clarification @iglovikov
\nYour quick and clear answers are highly appreciated.
So, uploading a CSV qualifies for a valid submission (valid in terms of qualification for prizes) and it's not a standard Code Competition.
\nedit: further clarification from host (see below): to be eligble for prize money, you need to provide the code.
I thought the submission requires the model (for rerunning) but not the actual CSV file. That is a lot different from what I thought before.
", "votes": null, "replies": [] }, { "id": 993597, "author_name": "sheriytm", "author_url": "", "post_date": "09/01/2020 03:24:31", "content": "@ilu000, I believe your code will still be needed for verification when you win a price or medal. @iglovikov may clarify this for us.
", "votes": null, "replies": [] }, { "id": 1007788, "author_name": "iglovikov", "author_url": "", "post_date": "09/12/2020 13:55:43", "content": "Yep. If the team is \"in the money\" and are not able to provide the code that reproduces your submission => team will be disqualified.
\nIt is not the case for medal solutions. We will not check solutions that are \"in medals\" but not \"in the money\".
\nIn the last year Lyft's competition second-place team refused to provide the code => they were kicked from the LB. The person who was at the 4th place become third and he provided the code.
", "votes": null, "replies": [] }, { "id": 1008240, "author_name": "sheriytm", "author_url": "", "post_date": "09/12/2020 21:43:49", "content": "Thanks for answering @iglovikov.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "988138": "Does everything (process/train/valid/predict) have to be done in the 9 hour kernel? \n\nOr can people train offline and import weights ?", "988147": "I believe you can train offline, but include the seperate notebook in the current one!", "988336": "jmaslek The requirement is that GPU notebook should have run time less than 9 hours and you can use TPU for training and TPU notebook should have runtime less than 3 hours. You can train and import weights. This is my understanding. Probably you can also weight for host comment on this.", "988654": "I would love to hear a word from the hosts about submissions such as: https://www.kaggle.com/corochann/save-your-time-submit-without-kernel-inference\n\nIt's probably not allowed as a valid submission but it raises the question of how much we are allowed to pre-compute.", "988952": "My understanding is that you could process/train/valid on your own machine and submit results by inference kernel on kaggle.", "989038": "That seems to make sense. I dont have any fancy offline GPUs to train on, so I was hoping I would have a shot to be competitive given the dataset size and time restrictions.", "989502": "You can train and perform inference offline and upload only the resulting CSV file to the inference kernel.", "989915": "Thank you very much for this clarification @iglovikov \nYour quick and clear answers are highly appreciated. \n\nSo, uploading a CSV qualifies for a valid submission (valid in terms of qualification for prizes) and it's not a standard Code Competition.\nedit: further clarification from host (see below): to be eligble for prize money, you need to provide the code.", "993513": "I thought the submission requires the model (for rerunning) but not the actual CSV file. That is a lot different from what I thought before.", "993597": "ilu000, I believe your code will still be needed for verification when you win a price or medal. @iglovikov may clarify this for us.", "1007788": "Yep. If the team is \"in the money\" and are not able to provide the code that reproduces your submission => team will be disqualified. \n\nIt is not the case for medal solutions. We will not check solutions that are \"in medals\" but not \"in the money\".\n\nIn the last year Lyft's competition second-place team refused to provide the code => they were kicked from the LB. The person who was at the 4th place become third and he provided the code.", "1008240": "Thanks for answering @iglovikov." }, "source": "meta" }